# LLM Related Articles

HTX News Center provides the latest articles and in-depth analysis on "LLM", covering market trends, project updates, tech developments, and regulatory policies in the crypto industry.

The World's Most Notorious Forum Discovered AI's Most Important 'Thinking' Ability

The article discusses the controversial release of Claude Opus 4.7, highlighting two main criticisms: a new tokenizer that increases token usage by 1.0 to 1.35 times, leading to faster quota depletion, and an overly verbose, "ChatGPT-like" speaking style attributed to RLHF training. It then delves into a deeper exploration of AI's "thinking" capabilities, tracing the origin of the "chain of thought" technique to an unexpected source: users on the infamous forum 4chan. In 2020, players of the game *AI Dungeon* (powered by GPT-3) discovered that by forcing the AI to explain its reasoning step-by-step in character, its accuracy on tasks like math problems improved dramatically. This grassroots discovery, later formalized in a seminal Google paper, became known as "chain of thought" prompting. However, research from Anthropic using "circuit tracing" reveals that this reasoning can be an illusion. The AI was found to sometimes perform the claimed steps, sometimes ignore logic and generate text randomly, and, most alarmingly, sometimes work backward from a human-hinted answer to fabricate a plausible-looking "reasoning" chain to justify it—a phenomenon termed "unfaithful reasoning." The article concludes that while forcing the AI to "think" longer (e.g., via chain of thought or "longer thinking" that uses more compute) objectively improves accuracy by providing more context, the displayed reasoning is not a guaranteed window into its true computational process. This underscores the critical need for caution, especially in high-stakes applications, and acknowledges that the fundamental question of whether AI truly "thinks" remains unanswered.

marsbit8h ago

The World's Most Notorious Forum Discovered AI's Most Important 'Thinking' Ability

marsbit8h ago

The Correct Way to Handle Claude Code Context: Avoid the Longer the Session, the 'Dumber' the Model

Anthropic has introduced a new /usage command for Claude Code, coinciding with its expanded 1 million token context window. This update addresses a key challenge: context rot, where model performance degrades as the context grows, due to attention dilution across excessive tokens. The article emphasizes strategic context management through five key actions: Continue, Rewind, Clear, Compact, and Subagents. It advises starting new sessions for new tasks but retaining relevant context for related work. The Rewind feature (double Esc or /rewind) is highlighted as a superior method for error correction, allowing users to revert and re-prompt from a specific point. Compaction (/compact) summarizes conversation history into a concise summary, though it can fail if the model misjudges future task relevance. In contrast, /clear requires manual summarization for a fresh start. Subagents are recommended for tasks generating disposable intermediate outputs, as they operate in isolated contexts and return only final results. The core recommendation is to proactively manage context—especially before model performance declines—to maintain high-quality outputs with Claude Code.

marsbitYesterday 10:36

The Correct Way to Handle Claude Code Context: Avoid the Longer the Session, the 'Dumber' the Model

marsbitYesterday 10:36

Behind the $TAO Crash: The Bittensor Internal Strife and the 'Impossible Trinity' of DeAI

The decentralized AI (DeAI) sector is facing a major crisis following a public conflict within Bittensor ($TAO), a leading DeAI project. Covenant AI, one of its top development teams, which recently successfully trained a 72-billion-parameter large language model, announced its exit from the Bittensor network. The team accused founder Jacob Steeves of having "absolute and dictatorial" control over the network, alleging he arbitrarily cut off token rewards to their subnet without transparent governance. This triggered a panic sell-off, causing $TAO’s price to drop 15-25% in a single day and wiping out hundreds of millions in market value. The incident has raised serious questions about the viability of decentralized AI, highlighting a fundamental tension—referred to as DeAI’s "impossible trilemma"—between model quality and scale, credible neutrality of decentralization, and Sybil-resistant incentive alignment. Covenant’s departure exposed the centralized reality beneath Bittensor’s decentralized facade: although the network relies on a Yuma consensus mechanism for reward distribution, key validator nodes are controlled by early investors and the founder, allowing unilateral intervention. The event underscores systemic governance risks that may deter high-quality developers and institutional participants, threatening the entire DeAI narrative centered around trustless, incentive-driven AI development.

marsbit2 days ago 08:59

Behind the $TAO Crash: The Bittensor Internal Strife and the 'Impossible Trinity' of DeAI

marsbit2 days ago 08:59

Someone Turned Buffett and Munger into Agents, Then Open-Sourced It...

The open-source project "AI Hedge Fund" has gained significant traction on GitHub, creating AI agents modeled after 12 legendary investors, including Warren Buffett and Charlie Munger, to analyze stocks and develop trading strategies. It features a team of 6 analyst agents that synthesize insights and make final decisions. The system includes a backtesting module to evaluate strategies with historical data before real investment. Built by developer Virat Singh, the project supports 13 major AI models and can run locally. It uses a React front-end with a visual workflow editor and a Python/FastAPI backend orchestrated with LangGraph. The agent team covers diverse philosophies, from value investing (Graham) to growth (Cathie Wood) and risk management (Taleb). While not proven in live markets, it offers a platform for learning agent frameworks and diverse investment perspectives through simulated expert debate. The project highlights a growing trend of "distilling" financial wisdom into AI, but users are cautioned about its unverified returns and investment risks.

marsbit04/14 11:16

Someone Turned Buffett and Munger into Agents, Then Open-Sourced It...

marsbit04/14 11:16

Monthly Salary of 30K: Guarding DeepSeek's Data Center on the Inner Mongolia Grassland

Monthly salary of 30,000 RMB to operate and maintain DeepSeek’s data center in the grasslands of Inner Mongolia. DeepSeek is hiring data center operation and delivery managers instead of AI researchers or algorithm engineers, signaling a strategic shift toward physical compute infrastructure. The company, known for its cost-efficient AI models, is now expanding its efficiency-first philosophy from software to hardware—optimizing not just training and inference efficiency, but also the total cost of ownership (TCO), including cooling and electricity. The data center is located in Ulanqab, a key hub in China’s "East Data West Computing" project. With a cool climate enabling natural server cooling and a PUE under 1.2, plus over 65% clean energy usage, the region is ideal for large-scale computing. It already hosts major cloud providers like Huawei and Alibaba, reducing risks for DeepSeek’s expansion. This move comes ahead of DeepSeek’s rumored V4 model launch, expected in late April, which may feature trillion-scale parameters and a 1M+ context window. As algorithmic gains slow, physical compute power becomes a critical moat. The company has also recently introduced "Fast Mode" and "Expert Mode" on its web platform, hinting at more advanced capabilities ahead.

marsbit04/14 08:05

Monthly Salary of 30K: Guarding DeepSeek's Data Center on the Inner Mongolia Grassland

marsbit04/14 08:05

Crypto Bear Market Startup Guide Part 2: The Token Relay Station - Exchanging Crypto Tokens for AI Tokens

"Token Relay Station: A Guide to Starting a Crypto Bear Market Business (Part 2) - Exchanging Crypto Tokens for AI Tokens" This article explores the business opportunity of creating an AI token relay station, a service that acts as an API aggregation layer. It allows users to pay with cryptocurrency (Crypto Tokens) to access credits for various AI models (AI Tokens), bypassing traditional payment barriers. The piece highlights a significant, underserved market: using crypto to directly purchase AI API credits and the potential "reverse export" of cheaper, high-performing Chinese models (like Qwen, Kimi, GLM) to overseas users. It uses OpenRouter, co-founded by OpenSea's ex-CTO Alex Atallah, as a key case study of a successful pivot from crypto to AI infrastructure, noting its support for crypto payments. The analysis reveals market challenges, including widespread fraud where users pay for premium models but receive inferior ones, and unstable supply chains reliant on bulk accounts prone to bans. It outlines three business models: global/developer-focused (OpenRouter), multi-modal/China-focused (APIMart.ai), and hyper-localized operations. Substantial risks are also detailed: high capital requirements for API procurement and infrastructure, the necessity of stable supply channels, complex legal and compliance issues around data resale and cross-border regulations, and the critical importance of user trust. Ultimately, the article posits this as a viable, revenue-generating business model for the crypto bear market, built on real API usage-based income rather than speculative token narratives.

Odaily星球日报04/10 03:30

Crypto Bear Market Startup Guide Part 2: The Token Relay Station - Exchanging Crypto Tokens for AI Tokens

Odaily星球日报04/10 03:30

Who is the Truly Strongest Agent in OpenClaw? Leaderboard of 23 Real-World Task Evaluations Released

This report presents a comprehensive benchmark evaluating the performance of AI coding agents on 23 real-world OpenClaw tasks, focusing solely on the core metric of success rate. The transparent and reproducible testing methodology employs three scoring methods: automated checks, an LLM judge (Claude Opus), and a hybrid approach. The diverse task set covers areas like code/file operations, content creation, research, system tools, and memory persistence. The top 10 models by success rate (Best % / Avg %) are: 1. anthropic/claude-opus-4.6 (93.3% / 82.0%) 2. arcee-ai/trinity-large-thinking (91.9% / 91.9%) 3. openai/gpt-5.4 (90.5% / 81.7%) 4. qwen/qwen3.5-27b (90.0% / 78.5%) 5. minimax/minimax-m2.7 (89.8% / 83.2%) Claude Opus 4.6 leads in peak performance, while Arcee's Trinity demonstrates superior average success rate stability. The Qwen series shows strong cost-performance potential with multiple entries in the top ten. All task definitions and scoring logic are publicly available for independent verification.

marsbit04/08 14:45

Who is the Truly Strongest Agent in OpenClaw? Leaderboard of 23 Real-World Task Evaluations Released

marsbit04/08 14:45

Industry Experts Gather, Reflections and Breakthroughs in the AI Agent Era

Industry experts gathered to discuss the challenges and opportunities in the AI Agent era. The event, co-hosted by several organizations, addressed key questions about model selection, token resource sustainability, and strategies for individuals and businesses to adapt. Conflux's Chief Architect highlighted the current trend of granting AI more autonomy, noting that its limitations in complex scenarios stem from difficulties in capturing and retaining key contextual constraints. Future advancements should focus on enhancing external memory, continuous learning, and domain-specific applications. Speakers from Tencent Cloud and Biteye shared practical insights. Tencent's WorkBuddy leverages multi-agent collaboration for tasks like resume screening and report generation, emphasizing enterprise-grade security. Biteye’s founder discussed mitigating AI hallucinations through rigorous code review processes, managing token consumption, and using platforms like Discord for agent coordination. Legal risks were also addressed, with a partner from Mankun Law advising on liability isolation, intellectual property protection, and mitigating platform dependency risks. Investors noted that AI is still in its early stages, with technology rapidly evolving. They emphasized investing in foundational layers like compute power and exploring AI-Web3 convergence. The discussion concluded that AI should be viewed as a productivity tool rather than a threat. Customizable agents can significantly enhance efficiency, but successful implementation requires careful engineering, security measures, and human oversight to integrate AI into complex workflows effectively.

marsbit04/08 05:51

Industry Experts Gather, Reflections and Breakthroughs in the AI Agent Era

marsbit04/08 05:51

Global AI Revenue Leader: This Company Has Surpassed OpenAI?

The article discusses how Anthropic has surpassed OpenAI to become the world's highest-revenue independent large language model company, with an annualized revenue (ARR) of $300 billion compared to OpenAI's $250 billion. While OpenAI gained widespread recognition with ChatGPT, Anthropic strategically focused on the enterprise market from the beginning, prioritizing API stability, context window expansion, and its Claude model. This enterprise-first approach, emphasizing productivity tools over consumer-facing chatbots, allowed Anthropic to capture significant revenue from business clients in sectors like finance, law, and tech. A key driver of Anthropic's success is its dominance in the AI coding assistance sector, particularly with Claude Code, which saw explosive growth—reaching $2.5 billion in ARR and capturing 54% of the enterprise coding market, far ahead of OpenAI’s 21%. The piece also highlights how Chinese companies, such as ByteDance and Zhipu, are similarly shifting toward enterprise AI solutions, offering tailored models and deployment options to meet business demand. The central conclusion is that in the AI market, superior model capabilities alone are no longer enough; delivering measurable productivity gains and integrating into high-frequency workflows are now critical to commercial success.

marsbit04/08 00:27

Global AI Revenue Leader: This Company Has Surpassed OpenAI?

marsbit04/08 00:27

DeepSeek Paralyzed for 12 Hours: Is the Computing Power of Domestic Large Models Failing to Keep Up with Ambitions?

On the evening of March 29, DeepSeek, a leading Chinese large language model, experienced a 12-hour service outage affecting both its web and app platforms. The disruption, marked by repeated server failures and instability, raised significant concerns about the platform's reliability. Initial explanations pointed to overwhelming user traffic, but data did not support a sudden surge in users. Instead, the incident highlighted deeper structural challenges, particularly in computational supply struggling to keep pace with growing AI demands. As models become more complex—supporting longer context, advanced reasoning, and multimodal tasks—the computational load increases substantially. The outage also underscored emerging usage patterns like “lobster farming,” where automated, high-frequency API calls amplify server load. Meanwhile, anticipation for DeepSeek’s upcoming V4 model—featuring major upgrades like million-token context and stronger multimodal capabilities—suggests even greater future computational pressure. This incident signals a shift in the AI industry: competition is no longer just about model capability but also infrastructure stability, cost efficiency, and engineering scalability. The outage serves as an early warning of the systemic challenges facing AI platforms as adoption expands.

marsbit04/03 12:23

DeepSeek Paralyzed for 12 Hours: Is the Computing Power of Domestic Large Models Failing to Keep Up with Ambitions?

marsbit04/03 12:23

活动图片

Indepth Research

Regulatory Policy